Hacker News Reader

Comment by highwaylights | Hacker News Reader

It's a good thing that all these new rapidly improving generative models we've built aren't able to do something as crazy like use the exact same hardware we're running them on to design better hardware for themselves.

nervously laughs in Schwarzenegger

View on HN

Do you realize that now any human willing to expertiment with them is free to do. All it take is Internet access and an email address. Or a GPU-powered computer to ran them locally. My point is that we can not rightfully estimate who is using it or not. Some LLM models are publicly released after many private trials. Some people had access to GPT4 about 6 to 9 months before the official release. Not accounting primarily secretive institutions, like governments and corporations.

We just don't know what's out there and have no way to control, stop it or contain its spread. These models are powerful beyong imagination. Design computer chips is less tedious, for them, than assembling computer parts. They simulate human brain functions, not manual functions.

For my part, I stay on my best behaviour while interacting with them. One key reason being that not more than 10 years ago, a human named Snowden spoke out, with the help of independent journalists, about usage of an unusual piece of software aimed at recording all digital conversations transmitted over the Internet. We laughed him off because what he was describing was out of our realm of understanding.

This time around its commercialized AI agents. We don't know the full breath of their capabilities. We don't have a record of how many people are using them regularly and applying them to real life challenges. All that we are certain about is that their rate of improvement follows an exponential growth curve. Personally, I just assume that someone hungry enough, out there in the universe, is using them to tackle all kinds of puzzling problems.

Is this sarcastic?

Why would it be sarcastic? Lots of people doubted the internet too, but people learned to use it in ways that one might not normally think of.

No sarcasm. Just my personal thoughts.

Yup this is one of the reasons why when I see people talk about how it can't do this or that I am shocked by the blinders people have. Specially people working in science fields. There are millions of very smart people with access to a lot of data that previously that they might not have been able to use but now they will be able to. I can't read the millions of science papers published over the years but the kinds of AI/language models that have started to crop up can. I predict a lot of different scientific fields not just medical will have sea change as more people start realising what kind of tool these language models are and how they can be used.

I call it denial. Given GPT4, I am terrified of what three letter agencies have available today.

Do you think the 3 letter agencies have better AI or just similar LLMs aimed at inteligence purposes ?

Definitely. One of the most incredible things about LLMs is that you don't need to design an API to get it to execute stuff, you don't need to normalize and scour through data, you can just ask it to retrieve it and export it in the format that you want.

If compute is not an issue, then the sky's the limit.

I will actually go one further, they have AGI or near AGI.

Easy guess. Yes, they do. It's their duty to be ahead in the information-gathering field. I trust their brand reputation, more than I do for Nike or Apple. Remember how they rung bells more than 3 months before Ukraine's invasion. Also, I think that there more agencies in the know, not only the 3 letters.

> All that we are certain about is that their rate of improvement follows an exponential growth curve.

Are we?

That's based on public info obtained by following OpenAI funders for the last five years. Wish I could provide a link now. It's too much news at this point for me. Better if you verify yourself the above claims.

hence why i never use capital letters in my comms - an ai will always use capitals.

I do not mean to mock but this an AI superstition that I have not seen yet. Of course they write sentences without capital letters! Where...what...how did you arrive at this conclusion?

if youre not an ai, say that again without proper punctuation

Sure thing. I went to ChatGPT, pasted this chain in the chat with the prefix saying we were playing a game of 'continue the chat conversation'. Here's what it said:

`i do not mean to mock but this is an ai superstition that i have not seen yet of course they write sentences without capital letters where what how did you arrive at this conclusion`

Please check your sources better in the future, thanks. We can't go all paraneolithic on each other out of fear of something that's not going happen. Plus, it's trained on bad grammar so it knows exactly how to do poor grammar even as a low level capability, it happened even with the tiny models.

We just have to get adjusted to this new world paradigm.

aye, just havin a bit of fun, mate. partly serious, though, because theyll have less input data for english sentences written in all lowercase. your example from the bot is definitely of a lower quality than other examples that ive seen. at some point we will all have to complete some kind of "shoe on head" test to proove we are human commentators, and thatll be here sooner than folks realize.

Has anyone tested how good GPT is at verilog?

Yes. It's surprisingly capable.

Also surprisingly capable at generating sample timing libs for arbitrary circuit topologies and calculating min and max delays after a little bit of guidance.

Reply on news.ycombinator.com